Intelligent sensor network multi-task deployment drift detection method and device

Abstract

This application discloses a method and apparatus for detecting multi-task deployment drift in an intelligent sensor network, comprising: collecting multi-task monitoring data of sensor nodes in the intelligent sensor network and generating node feature tensors based on the multi-task monitoring data; constructing a dynamic adjacency matrix for monitoring tasks; weighting and fusing the dynamic adjacency matrix using learnable task weights to generate a shared adjacency matrix; inputting the node feature tensors and the shared adjacency matrix into a spatiotemporal graph convolutional network and outputting M prediction results in parallel; calculating the prediction error based on the prediction results and actual monitoring data, and normalizing the prediction error; calculating the accuracy weight and detection weight of each monitoring task based on the prediction error, and weighting and averaging the accuracy weight and detection weight to obtain the fusion weight of each monitoring task; weighted summing of the normalized prediction errors to obtain the fusion drift score; and outputting a deployment drift decision signal when the drift score is greater than the dynamic detection threshold.

Description

Technical Field

[0001] This application relates to the field of sensor network management, and in particular to a method and apparatus for drift detection in multi-task deployment of intelligent sensor networks. Background Technology

[0002] Intelligent sensor networks, comprised of numerous autonomous devices deployed in a distributed and self-organizing manner, are primarily responsible for sensing the environment and collecting data, and have been widely used in various fields such as agriculture, forestry, and cities. However, in real-world scenarios, even after deployment, sensor networks can still be affected by environmental changes. Over time, the monitoring effectiveness of deployed sensor networks can be impacted by uncontrollable factors (such as sensor lifecycles and target changes), leading to time-drift characteristics in the collected data. This drift can weaken network coverage and connectivity between sensor nodes. Current research largely focuses on optimizing intelligent sensor network deployment by dynamically adjusting sensor node positions to respond to changes in monitored targets. Research involving adjusting the initially determined sensor node layout based on the latest real-time target information still prioritizes achieving optimal sensor network deployment. However, research gaps remain regarding the detection of time-drift in sensor network deployment and the determination of whether the sensor network needs updating.

[0003] Existing sensor network drift detection methods are mainly based on single-task identification and focus on single-task parameters. When a sensor network identifies and monitors multiple tasks with different parameters, it cannot comprehensively evaluate the overall deployment status of the sensor network for multiple tasks. When the monitoring data of a single task drifts, the detection system struggles to determine whether redeploying the sensor network is a better choice for the overall situation, and whether redeploying the sensor network will reduce the accuracy of identification for other tasks.

[0004] Regarding the problem that existing sensor network drift detection methods cannot comprehensively assess the overall deployment status of sensor networks across multiple tasks and are difficult to accurately determine when sensor networks need to be redeployed, no effective solution has yet been proposed. Summary of the Invention

[0005] To address this, this application provides a method and apparatus for detecting drift in multi-task deployment of intelligent sensor networks, thereby overcoming the problem that existing sensor network drift detection methods cannot comprehensively assess the overall deployment status of the sensor network for multiple tasks and are difficult to accurately determine when the sensor network needs to be redeployed.

[0006] To achieve the above objectives, according to a first aspect of the embodiments of this application, a method for multi-task deployment drift detection of an intelligent sensor network based on a spatiotemporal graph neural network is provided, comprising: collecting multi-task monitoring data from N sensor nodes in the intelligent sensor network, and generating a node feature tensor for each sensor node based on the N multi-task monitoring data, wherein the multi-task monitoring data includes environmental monitoring data collected by the sensor nodes performing M monitoring tasks, the types of environmental monitoring data included in different monitoring tasks are different, and M and N are both positive integers; constructing a dynamic adjacency matrix for each monitoring task based on the similarity between the geographic coordinate data of the N sensor nodes and the N multi-task monitoring data, obtaining M dynamic adjacency matrices; weighting and fusing the dynamic adjacency matrices of the M monitoring tasks through learnable task weights to generate a shared adjacency matrix; and combining the N node feature tensors with the learned task weights. The shared adjacency matrix input spatiotemporal graph convolutional network extracts spatiotemporal features through at least one shared spatiotemporal convolutional block and outputs the prediction results of the M monitoring tasks in parallel. Based on the M prediction results and the actual monitoring data of the M monitoring tasks, the prediction errors of the M monitoring tasks are calculated respectively, and the M prediction errors are normalized to obtain M normalized prediction errors. Based on the M prediction errors, the accuracy weight and detection weight of each monitoring task are calculated respectively, and the accuracy weight and detection weight are weighted and averaged to obtain the fusion weight of each monitoring task. Based on the M fusion weights, the M normalized prediction errors are weighted and summed to obtain a fusion drift score. If the fusion drift score is greater than the dynamic detection threshold, a deployment drift decision signal is output, wherein the deployment drift decision signal is used to indicate the redeployment of the intelligent sensor network.

[0007] In an optional embodiment, a dynamic adjacency matrix is ​​constructed for each monitoring task based on the similarity between the geographic coordinate data of the N sensor nodes and the N multi-task monitoring data. This includes: calculating a first geographic distance between any two sensor nodes based on the geographic coordinate data using spherical geometry formulas, filtering the obtained first geographic distances through a preset distance threshold to obtain multiple second geographic distances, and generating a geospatial constraint matrix for the N sensor nodes based on the multiple second geographic distances, wherein all the multiple second geographic distances are lower than the preset distance threshold; calculating a first similarity metric value between any two sensor nodes based on the L2 norm, filtering the obtained first similarity metric values ​​through a preset similarity threshold to obtain multiple second similarity metric values, and generating a feature similarity constraint matrix for the N sensor nodes based on the multiple second similarity metric values, wherein all the multiple second similarity metric values ​​are higher than the preset similarity threshold; and combining the geospatial constraint matrix with the feature similarity constraint matrix to generate the dynamic adjacency matrix, wherein the dynamic adjacency matrix is ​​a sparse adjacency matrix.

[0008] In an optional embodiment, a shared adjacency matrix is ​​generated by weighting and fusing the dynamic adjacency matrices of the M monitoring tasks using learnable task weights. This includes: setting trainable parameters for each monitoring task to obtain M trainable parameters; normalizing the M trainable parameters using a Softmax function to obtain weight coefficients for the M monitoring tasks, wherein the learnable task weights include the M weight coefficients, the sum of the M weight coefficients is a target value, and all M weight coefficients are non-negative; and weighting and fusing the dynamic adjacency matrices of the M monitoring tasks according to the M weight coefficients to generate the shared adjacency matrix.

[0009] In an optional embodiment, the shared spatiotemporal convolutional block includes: a first gated temporal convolutional layer, used to process N node feature tensors based on the shared adjacency matrix in spatial topology, extract the dynamic patterns of the N node feature tensors in the temporal dimension, and output a first temporal feature; a spatial graph convolutional layer, connected after the first gated temporal convolutional layer, used to perform graph convolution operations on the first temporal feature using the shared adjacency matrix as spatial constraints, and aggregate neighborhood information in the first temporal feature to generate a spatial association feature; and a second gated temporal convolutional layer, connected after the spatial graph convolutional layer, used to perform temporal convolution processing on the spatial association feature and output a second temporal feature, wherein the spatiotemporal feature is a comprehensive feature formed by cascading processing of the first temporal feature, the spatial association feature, and the second temporal feature.

[0010] In an optional embodiment, the prediction errors of the M monitoring tasks are calculated based on the M prediction results and the actual monitoring data of the M monitoring tasks, and the M prediction errors are normalized to obtain M normalized prediction errors, including: for monitoring task m among the M monitoring tasks, using the formula... Calculate the prediction bias of the i-th sensor node for the monitoring task m, where, Let be the prediction deviation of the i-th sensor for the monitoring task m. The prediction result of the i-th sensor for the monitoring task m. The actual monitoring data for the monitoring task m; expressed by the formula The prediction deviations of the N sensor nodes for the monitoring task m are spatiotemporally aggregated to obtain the prediction error of the monitoring task m, wherein, The prediction error of the monitoring task m; based on the historical data of the M monitoring tasks, using the formula The M prediction errors are normalized respectively, where, The normalized prediction error for the monitoring task m is... and These are the mean and standard deviation of the historical data for the monitoring task m, respectively.

[0011] In an optional embodiment, the precision weight and detection weight of each monitoring task are calculated based on the M prediction errors, and the precision weight and detection weight are weighted and averaged to obtain the fusion weight of each monitoring task. This includes: determining the precision weight of each monitoring task based on the current prediction errors of the M monitoring tasks; determining the detection weight of each monitoring task based on the historical detection performance indicators of the M monitoring tasks, wherein the historical detection performance indicators include at least precision and recall; and using hyperparameters... The fusion weight is obtained by weighting the accuracy weight and the detection weight, wherein the hyperparameters are... Used to indicate the importance of current performance.

[0012] In an optional embodiment, after weighted summation of the M normalized prediction errors based on the M fusion weights to obtain the fusion drift score, the method further includes: using the formula... Determine the dynamic detection threshold Where w is the length of the sliding window. This is the sensitivity coefficient. and These are the mean and standard deviation of the historical drift scores, respectively. A baseline level used to indicate the historical drift score. Used to indicate the fluctuation range of the historical drift score.

[0013] According to a second aspect of the embodiments of this application, a multi-task deployment drift detection device for intelligent sensor networks based on spatiotemporal graph neural networks is also provided, comprising: a data acquisition module, configured to acquire multi-task monitoring data of N sensor nodes in an intelligent sensor network, and generate a node feature tensor for each sensor node based on the N multi-task monitoring data, wherein the multi-task monitoring data includes environmental monitoring data collected by the sensor nodes performing M monitoring tasks, the types of environmental monitoring data included in different monitoring tasks are different, and M and N are both positive integers; a graph flow construction module, configured to construct a dynamic adjacency matrix for each monitoring task based on the similarity between the geographic coordinate data of the N sensor nodes and the N multi-task monitoring data, thereby obtaining M dynamic adjacency matrices; a topology fusion module, configured to perform weighted fusion of the dynamic adjacency matrices of the M monitoring tasks through learnable task weights to generate a shared adjacency matrix; and a spatiotemporal prediction module, configured to combine the N node feature tensors and the shared adjacency matrix into a shared adjacency matrix. A shared adjacency matrix input spatiotemporal graph convolutional network extracts spatiotemporal features through at least one shared spatiotemporal convolutional block and outputs the prediction results of the M monitoring tasks in parallel. An error calculation module calculates the prediction error of each of the M monitoring tasks based on the M prediction results and the actual monitoring data of the M monitoring tasks, and normalizes the M prediction errors to obtain M normalized prediction errors. A weight calculation module calculates the accuracy weight and detection weight of each monitoring task based on the M prediction errors, and performs a weighted average of the accuracy weight and detection weight to obtain the fusion weight of each monitoring task. A drift calculation module performs a weighted summation of the M normalized prediction errors based on the M fusion weights to obtain a fusion drift score. A drift determination module outputs a deployment drift decision signal when the fusion drift score is greater than a dynamic detection threshold, wherein the deployment drift decision signal is used to indicate the redeployment of the intelligent sensor network.

[0014] According to a third aspect of the embodiments of this application, an electronic device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to execute the drift detection method for multi-task deployment of intelligent sensor networks based on spatiotemporal graph neural networks as described in the first aspect through the computer program.

[0015] According to a fourth aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the multi-task deployment drift detection method for intelligent sensor networks based on spatiotemporal graph neural networks as described in the first aspect.

[0016] Compared with existing technologies, this application proposes a multi-task deployment drift detection mechanism for intelligent sensor networks based on spatiotemporal graph neural networks. This mechanism abstracts the sensor network into a dynamic graph structure, and by constructing a spatiotemporal topology graph flow, it fully utilizes the spatial correlation and temporal dependence in sensor data. It uses a multi-task prediction algorithm based on STGCN to process multiple monitoring tasks simultaneously, thereby more accurately identifying the true deployment drift. This solves the problem in related technologies that existing sensor network drift detection methods cannot comprehensively evaluate the overall deployment status of the sensor network for multiple tasks and are difficult to accurately determine when the sensor network needs to be redeployed.

[0017] It should be understood that the description in the Summary Section is not intended to limit the key or essential features of the embodiments of this application, nor is it intended to restrict the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0018] The above and other features, advantages, and aspects of the embodiments of this application will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of the application. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0019] Figure 1 This is a flowchart of an optional drift detection method for multi-task deployment of intelligent sensor networks based on spatiotemporal graph neural networks, according to an embodiment of this application.

[0020] Figure 2 This is an optional dynamic graph flow diagram of an embodiment of this application;

[0021] Figure 3 This is a schematic diagram of an optional ST-GCN-based multi-task prediction algorithm according to an embodiment of this application.

[0022] Figure 4 This is a schematic diagram of an optional multi-task deployment drift determination mechanism according to an embodiment of this application;

[0023] Figure 5 This is a flowchart illustrating the overall algorithm of an optional multi-task deployment drift detection algorithm based on a spatiotemporal graph neural network according to an embodiment of this application.

[0024] Figure 6 This is a structural block diagram of an optional intelligent sensor network multi-task deployment drift detection device based on a spatiotemporal graph neural network according to an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0027] To address the technical problems existing in related technologies, this embodiment provides a drift detection method for multi-task deployment of intelligent sensor networks based on spatiotemporal graph neural networks. Figure 1 This is a flowchart of a multi-task deployment drift detection method for intelligent sensor networks based on spatiotemporal graph neural networks, according to an embodiment of this application. The process includes the following steps:

[0028] Step S102: Collect multi-task monitoring data from N sensor nodes in the intelligent sensor network, and generate a node feature tensor for each sensor node based on the N multi-task monitoring data. The multi-task monitoring data includes environmental monitoring data collected by the sensor node when performing M monitoring tasks. Different monitoring tasks include different types of environmental monitoring data. M and N are both positive integers.

[0029] Step S104: Based on the similarity between the geographic coordinate data of the N sensor nodes and the N multi-task monitoring data, construct a dynamic adjacency matrix for each monitoring task to obtain M dynamic adjacency matrices.

[0030] Step S106: The dynamic adjacency matrices of the M monitoring tasks are weighted and fused using learnable task weights to generate a shared adjacency matrix.

[0031] Step S108: Input the N node feature tensors and the shared adjacency matrix into the spatiotemporal graph convolutional network, extract the spatiotemporal features through at least one shared spatiotemporal convolutional block, and output the prediction results of the M monitoring tasks in parallel.

[0032] Step S110: Calculate the prediction error of each of the M monitoring tasks based on the M prediction results and the actual monitoring data of the M monitoring tasks, and normalize the M prediction errors to obtain M normalized prediction errors.

[0033] Step S112: Calculate the accuracy weight and detection weight of each monitoring task based on the M prediction errors, and then perform a weighted average of the accuracy weight and detection weight to obtain the fusion weight of each monitoring task.

[0034] Step S114: Based on the M fusion weights, the M normalized prediction errors are weighted and summed to obtain the fusion drift score;

[0035] Step S116: If the fusion drift score is greater than the dynamic detection threshold, output a deployment drift decision signal, wherein the deployment drift decision signal is used to indicate the redeployment of the smart sensor network.

[0036] Through the above steps, this application proposes a multi-task deployment drift detection mechanism for intelligent sensor networks based on spatiotemporal graph neural networks. This mechanism abstracts the sensor network into a dynamic graph structure, and by constructing a spatiotemporal topology graph flow, it fully utilizes the spatial correlation and temporal dependence in sensor data. It uses a multi-task prediction algorithm based on STGCN to process multiple monitoring tasks simultaneously, thereby more accurately identifying the true deployment drift. This solves the problem in related technologies that existing sensor network drift detection methods cannot comprehensively evaluate the overall deployment status of the sensor network for multiple tasks and are difficult to accurately determine when the sensor network needs to be redeployed.

[0037] Optionally, based on the similarity between the geographic coordinate data of the N sensor nodes and the N multi-task monitoring data, a dynamic adjacency matrix is ​​constructed for each monitoring task, including: calculating a first geographic distance between any two sensor nodes based on the geographic coordinate data using spherical geometry formulas, filtering the obtained first geographic distances through a preset distance threshold to obtain multiple second geographic distances, and generating a geospatial constraint matrix for the N sensor nodes based on the multiple second geographic distances, wherein all the multiple second geographic distances are lower than the preset distance threshold; calculating a first similarity metric value between any two sensor nodes based on the L2 norm, filtering the obtained first similarity metric values ​​through a preset similarity threshold to obtain multiple second similarity metric values, and generating a feature similarity constraint matrix for the N sensor nodes based on the multiple second similarity metric values, wherein all the multiple second similarity metric values ​​are higher than the preset similarity threshold; and combining the geospatial constraint matrix with the feature similarity constraint matrix to generate the dynamic adjacency matrix, wherein the dynamic adjacency matrix is ​​a sparse adjacency matrix.

[0038] In this embodiment, based on graph theory and dynamic systems theory, a smart sensor network containing N nodes is formally defined as a dynamic graph flow G. ,in For the time domain, express A graph model constructed from data from each node of the sensor network.

[0039] Graphs at each moment Modeled as an undirected simple graph ,in, Let be a set of nodes, representing sensor nodes; Let be the set of edges, representing the connection relationships between nodes; Let be the weight function, representing the connection strength. The graph construction process is equivalent to its adjacency matrix. The generation process, the elements of this adjacency matrix Represents a node and At any moment The connection strength.

[0040] The graph structure is constructed based on a dual constraint criterion: geospatial constraint and feature similarity constraint. The geospatial constraint ensures that only physically reachable nodes can establish communication connections, which aligns with the actual deployment characteristics of sensor networks. The feature similarity constraint captures the correlation between node monitoring data, reflecting the spatial relevance of environmental phenomena.

[0041] Based on the actual deployment of sensor networks, geographical accessibility calculations, which calculate the distance between two points on a sphere, need to consider the influence of the Earth's curvature. This embodiment uses the Haversine formula: ,in, For the Earth's radius, and These represent latitude and longitude, respectively.

[0042] Feature similarity measurement using an L2 norm-based similarity function can effectively capture the non-linear relationships between node features: ,in The scale parameter controls the similarity decay rate.

[0043] By combining geographical constraints and feature similarity, we obtain the sparse adjacency matrix representation for each task: , among which, among which and The distance threshold (i.e., the aforementioned preset distance threshold) and similarity threshold (i.e., the aforementioned preset similarity threshold) are set based on the physical characteristics of the sensor. Therefore, the graph flow of each task can be represented as a sequence of adjacency matrices. Its structural diagram is as follows Figure 2 As shown in the figure. Where M is the number of tasks.

[0044] In an optional embodiment, this application proposes a multi-task prediction algorithm based on ST-GCN. The core idea of ​​this algorithm is to construct a dynamically fused graph structure and utilize a spatiotemporal graph convolutional network to simultaneously capture the evolutionary patterns in the temporal dimension and the correlation features in the spatial dimension, thereby achieving parallel prediction of multi-task monitoring data. The overall architecture of the algorithm is as follows: Figure 3 As shown.

[0045] This algorithm effectively improves prediction accuracy and the ability to capture complex spatiotemporal patterns through shared graph structure fusion, multi-level feature extraction using shared spatiotemporal convolutional blocks, and a multi-task parallel output architecture. Specifically, firstly, it adaptively integrates the spatial correlation information of each task through a dynamic topology graph fusion mechanism; secondly, it extracts both temporal and spatial features simultaneously using shared spatiotemporal convolutional blocks; and finally, it achieves independent prediction for each monitoring task through a multi-task parallel output architecture.

[0046] Based on the comparison between the predicted results and the actual monitoring data, a drift score representing the deployment status can be calculated. When the score exceeds the detection threshold, it is determined that the sensor network has experienced deployment drift, and the corresponding maintenance or redeployment process needs to be initiated.

[0047] Optionally, a shared adjacency matrix is ​​generated by weighting and fusing the dynamic adjacency matrices of the M monitoring tasks using learnable task weights. This includes: setting trainable parameters for each monitoring task to obtain M trainable parameters; normalizing the M trainable parameters using the Softmax function to obtain weight coefficients for the M monitoring tasks, wherein the learnable task weights include the M weight coefficients, the sum of the M weight coefficients is a target value, and all M weight coefficients are non-negative; and weighting and fusing the dynamic adjacency matrices of the M monitoring tasks according to the M weight coefficients to generate the shared adjacency matrix.

[0048] Multi-task monitoring data in sensor networks exhibit rich spatial correlations and task-related characteristics. Different monitoring tasks share key parameters (such as temperature, humidity, and illumination) that demonstrate both physical coupling and statistical correlation. To fully utilize this information, we propose a dynamic multi-task graph structure fusion mechanism.

[0049] The core idea of ​​this mechanism is to use fusion weights to fuse the topologies of each task through prediction result feedback. For each task... We construct a dynamic adjacency matrix Then, a shared adjacency matrix is ​​generated through weighted fusion: To ensure that all weights are non-negative and sum to 1 (i.e., the aforementioned target value), and to avoid certain tasks excessively dominating the fusion process, the weight coefficients are... Through trainable parameters After normalization using the Softmax function, the following is obtained: .

[0050] Trainable parameters This can be optimized using the gradient descent algorithm. During training, the weight parameters are adjusted using dual gradient signals. The first signal comes from the gradient of the prediction error for each task, and the second signal comes from the error gradient of the adjacency matrix predictor. This constrains the rationality of the weight allocation, prevents over-adjustment, and maintains the robustness of the topological representation.

[0051] Optionally, the shared spatiotemporal convolutional block includes: a first gated temporal convolutional layer, used to process N node feature tensors based on the shared adjacency matrix in spatial topology, extract the dynamic patterns of the N node feature tensors in the temporal dimension, and output a first temporal feature; a spatial graph convolutional layer, connected after the first gated temporal convolutional layer, used to perform graph convolution operations on the first temporal feature based on the shared adjacency matrix in spatial constraints, and aggregate neighborhood information in the first temporal feature to generate a spatial correlation feature; and a second gated temporal convolutional layer, connected after the spatial graph convolutional layer, used to perform temporal convolution processing on the spatial correlation feature and output a second temporal feature, wherein the spatiotemporal feature is a comprehensive feature formed by cascading processing of the first temporal feature, the spatial correlation feature, and the second temporal feature.

[0052] The shared spatiotemporal convolutional block (ST-Conv Block) is the core processing unit of the entire model, employing a cascaded "time-space-time" structure. The data flow of a complete ST-Conv Block can be formally represented as a composite function of three layers of operations: ,in, , and These represent the first-layer gated temporal convolution, the spatial graph convolution, and the second-layer gated temporal convolution, respectively. and The first The input and output of the blocks. By stacking multiple ST-Conv Blocks, deeper networks can be built to extract higher-level spatiotemporal features.

[0053] (1) Gated temporal convolutional layer: Temporal feature extraction adopts a causal convolution architecture to ensure that the model is at the same time step. The output depends strictly on current and historical information. This satisfies the causal constraints of the prediction task. In the specific implementation, a one-dimensional convolutional kernel is used to operate in the time dimension, and the sequence length is maintained through a left-padding strategy.

[0054] Given input tensor ,in For batch size, Input the number of channels. The time window length, The number of nodes. First, the number of channels is doubled using a causal convolutional layer: Then, the output will be... Divide into two parts along the channel dimension, which serve as the main feature paths respectively. and gated paths : .

[0055] To address the vanishing gradient problem during deep network training and enable the model to retain long-term memory, residual connections are introduced. First, an alignment function is used... (Typically a 1×1 convolution) Processing input , making its dimensions and Matching yields the residual terms. The final output is: ,in, For the Sigmoid function, This indicates element-wise multiplication.

[0056] (2) Spatial graph convolutional layer: Spatial feature extraction is based on spectral graph theory. Chebyshev polynomials are used to perform K-order approximation on graph convolution operations to reduce computational complexity and achieve local filtering.

[0057] Using a shared adjacency matrix and input features from multi-task fusion The spatial graph convolution operation is defined as follows: ,in, The Chebyshev coefficients are learnable. It is a k-th order Chebyshev polynomial.

[0058] In the implementation of this layer, residual connections are also introduced to stabilize the training of deep networks: ,in This is the dimension-aligned residual input.

[0059] Furthermore, the multi-task prediction output layer adopts a multi-head output. After feature extraction from multiple ST-Conv blocks, a temporal convolutional layer is connected for temporal compression, and causal convolution is used to compress the temporal dimension to the prediction length.

[0060] The fully connected layer expands the channel dimension to M×N, where M is the number of tasks and N is the number of nodes, allowing the model to simultaneously predict outputs for all tasks and all nodes. The output tensor is split into M independent outputs, each corresponding to a prediction result for a specific task: ,in, Features of the hidden layer and These are the weights and bias vectors of the fully connected layer.

[0061] The adjacency matrix predictor uses an LSTM architecture to capture the long-term dependencies of the adjacency matrix sequence, learn the evolution of the topology, and predict the future adjacency matrix. Using a recently constructed fusion graph structure As input, where This indicates the selected sequence length. The adjacency matrix predictor uses a two-layer LSTM architecture, which divides each graph structure... adjacency matrix Flattened into a vector Constructing batch input tensors By capturing long-term dependencies in adjacency matrix sequences through gating mechanisms, the evolutionary patterns of the topology are learned, and future adjacency matrices are predicted. : , , , .in Indicates a hidden state. Indicates memory state, This represents the linear transformation matrix of the output layer. This represents the output layer bias vector.

[0062] Furthermore, this application proposes a multi-task deployment drift detection mechanism, the core of which lies in identifying systematic deviations between multi-task monitoring data and short-term historical patterns. Based on the multi-task prediction results obtained from the aforementioned prediction algorithm, embodiments of this application propose a drift detection mechanism that integrates multi-dimensional information, such as... Figure 4 As shown, by using error aggregation, adaptive weight allocation, and dynamic threshold determination, accurate deployment status assessment is achieved, providing a basis for decision-making in the deployment and maintenance of sensor networks.

[0063] This detection mechanism achieves accurate drift identification through a three-layer processing approach: first, it eliminates scale differences between tasks through error calculation and normalization; then, it adaptively allocates task weights based on a dual accuracy-detection criterion; and finally, it employs a dynamic threshold strategy to achieve sensitive and robust drift determination. This design ensures that the detection system can both quickly respond to real deployment drift and effectively avoid false alarms.

[0064] Optionally, based on the M prediction results and the actual monitoring data of the M monitoring tasks, the prediction errors of the M monitoring tasks are calculated respectively, and the M prediction errors are normalized to obtain M normalized prediction errors, including: for monitoring task m among the M monitoring tasks, using the formula... Calculate the prediction bias of the i-th sensor node for the monitoring task m, where, Let be the prediction deviation of the i-th sensor for the monitoring task m. The prediction result of the i-th sensor for the monitoring task m. The actual monitoring data for the monitoring task m; expressed by the formula The prediction deviations of the N sensor nodes for the monitoring task m are spatiotemporally aggregated to obtain the prediction error of the monitoring task m, wherein, The prediction error of the monitoring task m; based on the historical data of the M monitoring tasks, using the formula The M prediction errors are normalized respectively, where, The normalized prediction error for the monitoring task m is... and These are the mean and standard deviation of the historical data for the monitoring task m, respectively.

[0065] For each monitoring task m, the model outputs a predicted value. Compared with the actual monitoring value The absolute error between them is calculated as follows: This absolute error reflects the prediction bias of a single node on a specific task.

[0066] The spatiotemporal aggregation error is obtained by averaging the errors of all nodes: This aggregation error (i.e. the prediction error mentioned above) eliminates random fluctuations at the node level and highlights changes at the task level.

[0067] To eliminate differences in error scales between different tasks, normalization is required. Statistical normalization based on historical data is employed: ,in and These are the mean and standard deviation of the error of task m during the training phase, respectively.

[0068] In an exemplary embodiment, the precision weight and detection weight of each monitoring task are calculated based on the M prediction errors, and the precision weight and detection weight are weighted and averaged to obtain the fusion weight of each monitoring task. This includes: determining the precision weight of each monitoring task based on the current prediction errors of the M monitoring tasks; determining the detection weight of each monitoring task based on the historical detection performance indicators of the M monitoring tasks, wherein the historical detection performance indicators include at least precision and recall; and using hyperparameters... The fusion weight is obtained by weighting the accuracy weight and the detection weight, wherein the hyperparameters are... Used to indicate the importance of current performance.

[0069] The task weight calculation employs a dual criterion: accuracy weight and detection weight. The accuracy weight is based on the current error level. This weight assigns greater importance to tasks with smaller errors, reflecting their current reliability.

[0070] The detection weights are based on historical performance: ,in The F1 score is calculated based on binary labels, which combines the precision and recall of previous drift detection in this task, reflecting the long-term reliability of the task.

[0071] The final fusion weight is a weighted average of the two: hyperparameters It balances the importance of current performance with historical performance. This is determined through grid search. .

[0072] Optionally, after obtaining the fusion drift score by weighted summation of the M normalized prediction errors based on the M fusion weights, the method further includes: using the formula... Determine the dynamic detection threshold Where w is the length of the sliding window. This is the sensitivity coefficient. and These are the mean and standard deviation of the historical drift scores, respectively. A baseline level used to indicate the historical drift score. Used to indicate the fluctuation range of the historical drift score.

[0073] Multi-task fusion drift score calculation is a weighted sum: The drift score integrates the performance of all tasks and reflects the overall state of the sensor network.

[0074] Dynamic threshold design is based on the principles of statistical process control. ,in The length of the sliding window. This is the sensitivity coefficient. and These are the mean and standard deviation of the historical drift scores, respectively. Reflecting the benchmark level, Characterizing the amplitude of fluctuations Control the detection sensitivity. This is achieved by adjusting... A certain value can strike a balance between detection sensitivity and false alarm rate.

[0075] In an optional embodiment, this application proposes a multi-task deployment drift detection mechanism based on a spatiotemporal graph neural network, which achieves deployment state evaluation through dynamic topology fusion and parallel prediction. The core algorithm flow is as follows: Figure 5 As shown, it mainly includes three key stages: dynamic graph flow construction, multi-task spatiotemporal prediction, and drift score fusion.

[0076] Given a set of sensor network nodes and a multi-task monitoring data sequence Set the sliding window length to The goal of the algorithm is to determine in Whether the sensor network needs to be redeployed is determined by outputting a deployment drift decision signal.

[0077] First, in the dynamic graph flow construction phase, for each time window, a node feature tensor for each sensor node is generated based on the multi-task monitoring data sequence. Then, based on the node's geographic coordinates and the similarity of the monitoring data, a dynamic adjacency matrix is ​​generated for each monitoring task. Secondly, in the multi-task topology fusion and spatiotemporal prediction stage, the topological structures of each task are fused using learnable task weights to generate a shared adjacency matrix. The node feature tensors and the fused adjacency matrix are then input into the spatiotemporal graph convolutional network STGCN, employing a cascaded "time-space-time" structure for multi-step prediction. Finally, in the drift detection and decision-making stage, the node-level absolute error of each task is calculated, and the task accuracy weights are integrated. and detection weights Calculate the fusion drift score and determine whether the sensor network has experienced deployment drift based on the drift score at each time point.

[0078] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0079] Embodiments of this application also provide a multi-task deployment drift detection device for intelligent sensor networks based on spatiotemporal graph neural networks, such as... Figure 6 As shown, the device includes:

[0080] The data acquisition module 61 is used to acquire multi-task monitoring data of N sensor nodes in the intelligent sensor network, and generate a node feature tensor for each sensor node based on the N multi-task monitoring data. The multi-task monitoring data includes environmental monitoring data collected by the sensor node when performing M monitoring tasks. The types of environmental monitoring data included in different monitoring tasks are different, and M and N are both positive integers.

[0081] The graph flow construction module 62 is used to construct a dynamic adjacency matrix for each monitoring task based on the similarity between the geographic coordinate data of the N sensor nodes and the N multi-task monitoring data, thereby obtaining M dynamic adjacency matrices.

[0082] The topology fusion module 63 is used to perform weighted fusion of the dynamic adjacency matrices of the M monitoring tasks through learnable task weights to generate a shared adjacency matrix.

[0083] The spatiotemporal prediction module 64 is used to input the N node feature tensors and the shared adjacency matrix into the spatiotemporal graph convolutional network, extract spatiotemporal features through at least one shared spatiotemporal convolutional block, and output the prediction results of the M monitoring tasks in parallel.

[0084] The error calculation module 65 is used to calculate the prediction error of the M monitoring tasks based on the M prediction results and the actual monitoring data of the M monitoring tasks, and to normalize the M prediction errors to obtain M normalized prediction errors.

[0085] The weight calculation module 66 calculates the accuracy weight and detection weight of each monitoring task based on the M prediction errors, and then performs a weighted average of the accuracy weight and detection weight to obtain the fusion weight of each monitoring task.

[0086] The drift calculation module 67 is used to perform a weighted summation of the M normalized prediction errors based on the M fusion weights to obtain a fusion drift score.

[0087] The drift determination module 68 is used to output a deployment drift decision signal when the fusion drift score is greater than the dynamic detection threshold, wherein the deployment drift decision signal is used to indicate the redeployment of the smart sensor network.

[0088] Using the aforementioned device, this application proposes a multi-task deployment drift detection mechanism for intelligent sensor networks based on spatiotemporal graph neural networks. This mechanism abstracts the sensor network into a dynamic graph structure, fully utilizes the spatial correlation and temporal dependence in sensor data by constructing a spatiotemporal topology graph flow, and uses a multi-task prediction algorithm based on STGCN to process multiple monitoring tasks simultaneously, thereby more accurately identifying true deployment drift. This solves the problem in related technologies that existing sensor network drift detection methods cannot comprehensively assess the overall deployment status of the sensor network for multiple tasks and are difficult to accurately determine when the sensor network needs to be redeployed.

[0089] Optionally, the above-mentioned graph flow construction module 62 is further configured to calculate the first geographic distance between any two sensor nodes based on the spherical geometry formula according to the geographic coordinate data, filter the obtained first geographic distances through a preset distance threshold to obtain a plurality of second geographic distances, and generate a geospatial constraint matrix of the N sensor nodes based on the plurality of second geographic distances, wherein the plurality of second geographic distances are all lower than the preset distance threshold; calculate the first similarity metric value of multi-task monitoring data between any two sensor nodes based on the L2 norm, filter the obtained first similarity metric values ​​through a preset similarity threshold to obtain a plurality of second similarity metric values, and generate a feature similarity constraint matrix of the N sensor nodes based on the plurality of second similarity metric values, wherein the plurality of second similarity metric values ​​are all higher than the preset similarity threshold; combine the geospatial constraint matrix and the feature similarity constraint matrix to generate the dynamic adjacency matrix, wherein the dynamic adjacency matrix is ​​a sparse adjacency matrix.

[0090] Optionally, the aforementioned topology fusion module 63 is further configured to set trainable parameters for each monitoring task to obtain M trainable parameters; normalize the M trainable parameters using the Softmax function to obtain weight coefficients for the M monitoring tasks, wherein the learnable task weights include the M weight coefficients, the sum of the M weight coefficients is the target value, and all M weight coefficients are non-negative; and perform weighted fusion of the dynamic adjacency matrices of the M monitoring tasks based on the M weight coefficients to generate the shared adjacency matrix.

[0091] Optionally, the aforementioned shared spatiotemporal convolutional block includes: a first gated temporal convolutional layer, used to process N node feature tensors based on the shared adjacency matrix in spatial topology, extract the dynamic patterns of the N node feature tensors in the temporal dimension, and output a first temporal feature; a spatial graph convolutional layer, connected after the first gated temporal convolutional layer, used to perform graph convolution operations on the first temporal feature using the shared adjacency matrix as spatial constraints, and aggregate neighborhood information in the first temporal feature to generate a spatial association feature; and a second gated temporal convolutional layer, connected after the spatial graph convolutional layer, used to perform temporal convolution processing on the spatial association feature and output a second temporal feature, wherein the spatiotemporal feature is a comprehensive feature formed by cascading processing of the first temporal feature, the spatial association feature, and the second temporal feature.

[0092] Optionally, the aforementioned error calculation module 65 is further used to calculate the error for monitoring task m out of the M monitoring tasks using the formula... Calculate the prediction bias of the i-th sensor node for the monitoring task m, where, Let be the prediction deviation of the i-th sensor for the monitoring task m. The prediction result of the i-th sensor for the monitoring task m. The actual monitoring data for the monitoring task m; expressed by the formula The prediction deviations of the N sensor nodes for the monitoring task m are spatiotemporally aggregated to obtain the prediction error of the monitoring task m, wherein, The prediction error of the monitoring task m; based on the historical data of the M monitoring tasks, using the formula The M prediction errors are normalized respectively, where, The normalized prediction error for the monitoring task m is... and These are the mean and standard deviation of the historical data for the monitoring task m, respectively.

[0093] Optionally, the aforementioned weight calculation module 66 is further configured to determine the precision weight of each monitoring task based on the current prediction error of the M monitoring tasks; and to determine the detection weight of each monitoring task based on the historical detection performance indicators of the M monitoring tasks, wherein the historical detection performance indicators include at least: precision and recall; and through hyperparameters. The fusion weight is obtained by weighting the accuracy weight and the detection weight, wherein the hyperparameters are... Used to indicate the importance of current performance.

[0094] Optionally, the drift calculation module 67 described above is also used to calculate using a formula Determine the dynamic detection threshold Where w is the length of the sliding window. This is the sensitivity coefficient. and These are the mean and standard deviation of the historical drift scores, respectively. A baseline level used to indicate the historical drift score. Used to indicate the fluctuation range of the historical drift score.

[0095] Embodiments of this application also provide a storage medium including a stored program, wherein the program, when executed, performs any of the methods described above. Optionally, in this embodiment, the storage medium may be configured to store program code for performing the steps of the methods described above.

[0096] Optionally, in this embodiment, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0097] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.

[0098] Optionally, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor and the input / output device is connected to the processor.

[0099] Embodiments of this application also provide a computer program product, including a non-volatile computer-readable storage medium storing the computer program product, wherein the computer program, when executed by a processor, implements the steps of the methods described in various embodiments of this application.

[0100] Optionally, in this embodiment, the computer program described above can be configured to implement the steps in any of the above method embodiments when executed by a processor.

[0101] Optionally, specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0102] The scope of this application is defined by the appended claims rather than the foregoing description, and is therefore intended to encompass all variations falling within the meaning and scope of equivalents of the claims. No reference numerals in the claims should be construed as limiting the scope of the claims. Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units or devices recited in a device claim may also be implemented by a single unit or device in software or hardware. Terms such as "first," "second," etc., are used only for distinguishing descriptions and do not indicate any particular order, nor should they be construed as indicating or implying relative importance.

[0103] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily made by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims, and the above embodiments should be regarded as exemplary and non-limiting.

Claims

1. A method for drift detection in multi-task deployment of intelligent sensor networks based on spatiotemporal graph neural networks, characterized in that, include: Collect multi-task monitoring data from N sensor nodes in an intelligent sensor network, and generate a node feature tensor for each sensor node based on the N multi-task monitoring data. The multi-task monitoring data includes environmental monitoring data collected by the sensor node when performing M monitoring tasks. Different monitoring tasks include different types of environmental monitoring data, and M and N are both positive integers. Based on the similarity between the geographic coordinate data of the N sensor nodes and the N multi-task monitoring data, a dynamic adjacency matrix is ​​constructed for each monitoring task, resulting in M ​​dynamic adjacency matrices. The dynamic adjacency matrices of the M monitoring tasks are weighted and fused using learnable task weights to generate a shared adjacency matrix. The N node feature tensors and the shared adjacency matrix are input into the spatiotemporal graph convolutional network. The spatiotemporal features are extracted through at least one shared spatiotemporal convolutional block, and the prediction results of the M monitoring tasks are output in parallel. The shared spatiotemporal convolutional block includes: a first gated temporal convolutional layer, used to process N node feature tensors based on the shared adjacency matrix in spatial topology, extracting the dynamic patterns of the N node feature tensors in the temporal dimension, and outputting a first temporal feature; a spatial graph convolutional layer, connected after the first gated temporal convolutional layer, used to perform graph convolution operations on the first temporal feature using the shared adjacency matrix as spatial constraints, aggregating neighborhood information in the first temporal feature to generate a spatial association feature; and a second gated temporal convolutional layer, connected after the spatial graph convolutional layer, used to perform temporal convolution processing on the spatial association feature, and output a second temporal feature, wherein the spatiotemporal feature is a comprehensive feature formed by cascading processing of the first temporal feature, the spatial association feature, and the second temporal feature; Based on the M prediction results and the actual monitoring data of the M monitoring tasks, the prediction error of the M monitoring tasks is calculated respectively, and the M prediction errors are normalized to obtain M normalized prediction errors. Calculate the precision weight and detection weight for each monitoring task based on the M prediction errors, and then perform a weighted average of the precision weight and detection weight to obtain the fusion weight for each monitoring task. This includes: determining the precision weight for each monitoring task based on the current prediction errors of the M monitoring tasks; determining the detection weight for each monitoring task based on the historical detection performance indicators of the M monitoring tasks, wherein the historical detection performance indicators include at least precision and recall; and using hyperparameters... The fusion weight is obtained by weighting the accuracy weight and the detection weight, wherein the hyperparameters are... Used to indicate the importance of current performance; The fusion drift score is obtained by weighting and summing the M normalized prediction errors based on the M fusion weights. If the fusion drift score is greater than the dynamic detection threshold, a deployment drift decision signal is output, wherein the deployment drift decision signal is used to indicate the redeployment of the smart sensor network.

2. The method according to claim 1, characterized in that, Based on the similarity between the geographic coordinate data of the N sensor nodes and the N multi-task monitoring data, a dynamic adjacency matrix is ​​constructed for each monitoring task, including: Based on the spherical geometry formula, the first geographic distance between any two sensor nodes is calculated according to the geographic coordinate data. Multiple first geographic distances are filtered by a preset distance threshold to obtain multiple second geographic distances. A geospatial constraint matrix of the N sensor nodes is generated based on the multiple second geographic distances, wherein all of the multiple second geographic distances are lower than the preset distance threshold. The first similarity measure of task monitoring data between any two sensor nodes is calculated based on the L2 norm. Multiple first similarity measures are then filtered through a preset similarity threshold to obtain multiple second similarity measures. A feature similarity constraint matrix of the N sensor nodes is then generated based on the multiple second similarity measures, wherein all of the multiple second similarity measures are higher than the preset similarity threshold. The geospatial constraint matrix and the feature similarity constraint matrix are combined to generate the dynamic adjacency matrix, wherein the dynamic adjacency matrix is ​​a sparse adjacency matrix.

3. The method according to claim 1, characterized in that, The dynamic adjacency matrices of the M monitoring tasks are weighted and fused using learnable task weights to generate a shared adjacency matrix, including: Trainable parameters are set for each monitoring task, resulting in M ​​trainable parameters; The M trainable parameters are normalized using the Softmax function to obtain the weight coefficients of the M monitoring tasks. The weights of the learnable tasks include the M weight coefficients, the sum of the M weight coefficients is the target value, and all M weight coefficients are non-negative. The shared adjacency matrix is ​​generated by weighting and fusing the dynamic adjacency matrices of the M monitoring tasks according to the M weight coefficients.

4. The method according to claim 1, characterized in that, Based on the M prediction results and the actual monitoring data of the M monitoring tasks, the prediction errors of the M monitoring tasks are calculated respectively, and the M prediction errors are normalized to obtain M normalized prediction errors, including: For monitoring task m among the M monitoring tasks, the formula is used. Calculate the prediction bias of the i-th sensor node for the monitoring task m, where, Let be the prediction deviation of the i-th sensor for the monitoring task m. The prediction result of the i-th sensor for the monitoring task m. The actual monitoring data for the monitoring task m; Through formula The prediction deviations of the N sensor nodes for the monitoring task m are spatiotemporally aggregated to obtain the prediction error of the monitoring task m, wherein, The prediction error of the monitoring task m; Based on the historical data of the M monitoring tasks, using the formula The M prediction errors are normalized respectively, where, The normalized prediction error for the monitoring task m is... and These are the mean and standard deviation of the historical data for the monitoring task m, respectively.

5. The method according to claim 1, characterized in that, After weighting and summing the M normalized prediction errors based on the M fusion weights to obtain the fusion drift score, the method further includes: Through formula Determine the dynamic detection threshold Where w is the length of the sliding window. This is the sensitivity coefficient. and These are the mean and standard deviation of the historical drift scores, respectively. A baseline level used to indicate the historical drift score. Used to indicate the fluctuation range of the historical drift score.

6. A multi-task deployment drift detection device for intelligent sensor networks based on spatiotemporal graph neural networks, characterized in that, include: The data acquisition module is used to acquire multi-task monitoring data from N sensor nodes in the intelligent sensor network, and generate a node feature tensor for each sensor node based on the N multi-task monitoring data. The multi-task monitoring data includes environmental monitoring data collected by the sensor node when performing M monitoring tasks. Different monitoring tasks include different types of environmental monitoring data, and M and N are both positive integers. The graph flow construction module is used to construct a dynamic adjacency matrix for each monitoring task based on the similarity between the geographic coordinate data of the N sensor nodes and the N multi-task monitoring data, thereby obtaining M dynamic adjacency matrices. The topology fusion module is used to perform weighted fusion of the dynamic adjacency matrices of the M monitoring tasks through learnable task weights to generate a shared adjacency matrix. A spatiotemporal prediction module is used to input N node feature tensors and the shared adjacency matrix into a spatiotemporal graph convolutional network, extract spatiotemporal features through at least one shared spatiotemporal convolutional block, and output the prediction results of the M monitoring tasks in parallel. The shared spatiotemporal convolutional block includes: a first gated temporal convolutional layer, used to process the N node feature tensors based on the shared adjacency matrix in spatial topology, extracting the dynamic patterns of the N node feature tensors in the temporal dimension, and outputting a first temporal feature; a spatial graph convolutional layer, connected after the first gated temporal convolutional layer, used to perform graph convolution operations on the first temporal feature using the shared adjacency matrix as spatial constraints, aggregating neighborhood information in the first temporal feature to generate a spatial correlation feature; and a second gated temporal convolutional layer, connected after the spatial graph convolutional layer, used to perform temporal convolution processing on the spatial correlation feature, outputting a second temporal feature. The spatiotemporal feature is a comprehensive feature formed by cascading the first temporal feature, the spatial correlation feature, and the second temporal feature. The error calculation module is used to calculate the prediction error of the M monitoring tasks based on the M prediction results and the actual monitoring data of the M monitoring tasks, and to normalize the M prediction errors to obtain M normalized prediction errors. The weight calculation module calculates the precision weight and detection weight for each monitoring task based on the M prediction errors, and then performs a weighted average of the precision weight and detection weight to obtain the fusion weight for each monitoring task. This includes: determining the precision weight for each monitoring task based on the current prediction errors of the M monitoring tasks; determining the detection weight for each monitoring task based on the historical detection performance indicators of the M monitoring tasks, wherein the historical detection performance indicators include at least precision and recall; and using hyperparameters... The fusion weight is obtained by weighting the accuracy weight and the detection weight, wherein the hyperparameters are... Used to indicate the importance of current performance; The drift calculation module is used to perform a weighted summation of the M normalized prediction errors based on the M fusion weights to obtain the fusion drift score. The drift determination module is used to output a deployment drift decision signal when the fusion drift score is greater than the dynamic detection threshold, wherein the deployment drift decision signal is used to indicate the redeployment of the smart sensor network.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program, when executed, performs the method of any one of claims 1 to 5.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 5.

Images